<p>The aim was to demonstrate a non-destructive approach for estimating the quality and freshness of whole Indian mackerel (<i>Rastrelliger kanagurta</i>) stored under iced conditions. The colour changes in the eye region of fish during iced storage indicate different stages of spoilage and edibility. These colour changes in fish eyes were interpreted using deep learning, with feature extraction performed by convolutional layers. Fresh samples were stored under iced conditions for 15 days, and images of fish eyes were captured daily to create a database for developing the convolutional neural networks (CNN) model. Simultaneously, fish samples were subjected to destructive analysis through K-value and psychrophilic count estimations to label the model predictions with various freshness levels. The acquired image dataset was augmented by varying hue, saturation, contrast, blur, brightness, reflection, and Gaussian noise. The augmented dataset was used for training, optimization, and regularization of the CNN model. The training, validation, and test accuracies of the model were 92.19%, 83.57%, and 83.91%, respectively. The results of the destructive quality analysis were used to establish the overall quality limits in days, i.e., “Extremely Fresh (1–4 days), Fresh (5–11 days), or Spoiled (beyond 11 days).” The validation of the developed approach was performed using random fish samples, predicting ‘days after fish caught,’ ‘remaining shelf life in days,’ and the degree of freshness of Indian mackerel. The framework achieved an accuracy of 90%, demonstrating the effectiveness of the proposed approach, with practical implications for the industry and consumers.</p>